What's new

Changelog

Every release of Fulcrum, in reverse chronological order. Download the latest binaries from the download page.

  1. v0.12.4

    2026-05-30Latest

    Reliable tool calling with Google Gemini & Anthropic

    Reliability

    • ▸Google Gemini and Anthropic models now work reliably with connected tools — multi-step tool conversations no longer error out partway through.
    • ▸Connecting an external tool server (ClickUp, filesystem, and other MCP servers) no longer breaks Gemini or Anthropic models.

  2. v0.12.0

    2026-05-25

    Google Gemini 3.5 Flash, plus a /apps command

    New

    • ▸Added Google Gemini 3.5 Flash as a selectable model.
    • ▸New /apps command — list the applications installed on your Mac.

    Improvements

    • ▸Smoother computer-use scrolling.

  3. v0.11.0

    2026-05-22

    Quality-of-life input upgrades

    Improvements

    • ▸The prompt auto-focuses so you can start typing immediately.
    • ▸Paste context tags directly into the prompt.
    • ▸Right-click to copy from the transcript.
    • ▸New computer-use tool for controlling the desktop.

  4. v0.10.0

    2026-05-15

    Connect external tools, voice input, and image vision

    New

    • ▸Connect external tool servers (MCP) — including ClickUp and the filesystem — directly to Fulcrum.
    • ▸Voice input — run /voice for hands-free transcription.
    • ▸Vision — drag an image into the prompt to use image-capable models.

    Reliability

    • ▸Stability fixes for the Fulcrum Hotkey companion app.

  5. v0.9.6

    2026-05-12

    MAGPiE tool-calling fix

    Model fixes

    • ▸Fixed MAGPiE model 502 errors — scx.ai's MAGPiE does not support tool calling. The agent now detects this and sends chat-only requests automatically.
    • ▸Updated MAGPiE capabilities in the model catalog: reasoning + JSON only (no tools).

  6. v0.9.3

    2026-05-10

    Model picker polish, slash commands work on first Enter

    Model picker

    • ▸scx.ai models show their origin clearly. Fulcrum Inference models (⚡ ultra-fast and 💪 frontier) are visually distinct.
    • ▸Fixed Qwen3-32B not resolving — model ID corrected for the upstream API.

    Faster command execution

    • ▸Select a command from the picker and hit Enter — it runs immediately. No more pressing Enter twice.
    • ▸Arrow keys navigate the command list in both the terminal and the Hotkey companion. Escape dismisses the picker.

  7. v0.9.0

    2026-05-10

    Fulcrum Hotkey app, parallel sandbox agents, rock-solid stability

    Fulcrum Hotkey — access Fulcrum from anywhere

    • ▸A proper macOS app that lives in your menu bar. Download the DMG, drag to Applications, and Ctrl+Space opens Fulcrum instantly from any app.
    • ▸Zero setup — auto-starts on login, self-updates, and downloads the CLI on first launch if you don't have it yet.

    Fulcrum Sandbox Agents — parallel work at scale

    • ▸Split large tasks into parallel sub-tasks with `/remote`. Each agent works in its own isolated cloud sandbox — cloning your repo, making changes, and pushing to its own branch.
    • ▸A review agent automatically merges all branches, runs your tests, and produces one clean PR. Failed agents retry automatically.
    • ▸Included in every plan — Starter gets 2 concurrent agents, Pro gets 10, Team gets 50.

    Stability

    • ▸Your terminal never breaks after a crash — no more `^M` characters or stuck cursor. Fulcrum now restores terminal state cleanly on every exit.

  8. v0.8.4

    2026-05-11

    Drop images into the conversation

    Image drag-and-drop

    • ▸Drag an image into Fulcrum and it shows up as a clean `[#image 1]` tag — not a messy file path. Drop multiple images and they number sequentially.
    • ▸When you send the message, Fulcrum sends the actual images to the vision model automatically. Just drop, type your question, and hit Enter.

  9. v0.8.0

    2026-05-10

    More models, automatic failover, model picker

    ⚡ Ultra-fast models

    • ▸Four new models running on dedicated hardware — up to 1,200 tokens per second. Look for the ⚡ icon in the model picker to find the fastest options for quick iterations.
    • ▸All models include vision support, tool use, and reasoning capabilities. Switch between them instantly with `/models`.

    Automatic failover — no more downtime

    • ▸If a model is slow or unavailable, Fulcrum automatically switches to a backup within 10 seconds. You keep working without interruption.

    Hotkey companion upgrades

    • ▸New model picker — type `/models`, use arrow keys, hit Enter. See all available models with speed indicators and context sizes.
    • ▸Type `/` to see all available commands with descriptions. Autocomplete filters as you type.
    • ▸Toggle between compact and full-screen mode with Cmd+M or the □ button.
    • ▸Clean cost summary after each response — see input, output, cached tokens, and cost at a glance.

  10. v0.6.0

    2026-05-09

    Smart quotas — cheaper models give you more tokens

    Cost-aware quotas

    • ▸Your token budget now adapts to the model you're using. Cheaper models give you more tokens for the same subscription price — you're rewarded for choosing efficient models.
    • ▸Daily usage cap prevents accidentally burning your monthly budget in one session. Check your usage anytime with `/usage`.

    Menu bar presence

    • ▸Fulcrum logo in your macOS menu bar with online/offline status, restart, and version info.

  11. v0.4.1

    2026-05-04

    See where your tokens went

    Usage breakdown

    • ▸New `/usage` command shows exactly where your tokens went — month-to-date and last 24 hours, broken down by model, with cache efficiency ratios.
    • ▸Command output is now easy to read — bright text on dark backgrounds instead of barely-visible dim grey.

  12. v0.4.0

    2026-05-04

    No more Gatekeeper warnings on macOS

    Signed and notarised

    • ▸Every macOS binary is now signed with our Apple Developer certificate and notarised. Download, run, done — no security warnings, no terminal workarounds.

  13. v0.3.27

    2026-05-04

    Cleaner conversation view

    Less noise, more signal

    • ▸Tool calls that fail due to model mistakes are now handled silently — the model retries with the right tool automatically. Your transcript only shows real, successful operations.
    • ▸Tighter spacing between tool-call groups means you can see more of your conversation on screen.

  14. v0.3.26

    2026-05-04

    Faster responses after a pause

    Background cache warming

    • ▸Step away for a coffee and your next message is still fast. Fulcrum keeps the prompt cache warm in the background while you're thinking, so you never pay the cold-start penalty.
    • ▸Cache warming doesn't count against your quota — it's on us.

  15. v0.3.24

    2026-05-03

    Longer responses for complex tasks

    Extended output

    • ▸Frontier models can now generate up to 65,536 tokens per response — long planning sessions, detailed code reviews, and large refactors no longer get cut short.

  16. v0.3.23

    2026-05-02

    Slash command picker improvements

    Slash picker

    • ▸Two earlier patches helped but didn't fully fix the case where typing `/m` then `c` made the picker vanish instead of narrowing to `/mcp`. The picker now sits inside the prompt area and stays visible whenever there's at least one match, including when the list shrinks to a single row.

  17. v0.3.22

    2026-05-01

    Cap-reached menu + slash picker fix

    Over-cap UX

    • ▸Mid-session cap-reached errors used to dump a raw error blob into the transcript. You now get a clean arrow-key menu with four options — T to top up, U to upgrade, W to wait until reset, Q to quit. T and U open the right fulcrumcode.app page in your browser; W keeps your existing access working until your monthly period resets.

    Slash picker

    • ▸Fixed a layout bug where the slash picker disappeared when narrowing from many matches to one — e.g. typing `/m` then `c` made the picker vanish instead of showing only `/mcp`.

  18. v0.3.21

    2026-05-01

    Under the hood — stronger foundations, fewer silent failures

    Reliability & Diagnostics

    • ▸Fulcrum used to silently swallow errors in places you'd never notice — like when cost tracking failed, or a memory file was corrupted, or the keychain was unreachable. Now every one of these produces a logged diagnostic so our team can identify and fix issues before you even see them.
    • ▸Stream interruptions (network hiccups, provider blips) were already retried automatically — but now the exact reason for each retry is recorded, making it much faster to diagnose intermittent connectivity problems.

    Code Quality

    • ▸A major internal clean-up removed over 900 lines of duplicated code — the kind of duplication that leads to bugs being fixed in one place but not another. Skills, agents, and your memory store now all share the same well-tested foundations.
    • ▸The tool-output renderer that powers your transcript (diffs, bash output, search results) was restructured from one large file into focused, single-responsibility modules. This makes future improvements faster and less likely to introduce regressions.

    Type Safety

    • ▸Stricter type-checking is now enforced across the codebase, catching a class of latent bugs before they ever ship.

  19. v0.3.20

    2026-05-01

    Tighter transcript — no more huge gaps between tool blocks

    Transcript

    • ▸Killed the implicit blank line above every tool block. With non-groupable tools (Edit/Bash/Task) sitting between read-only Searched groups, the cumulative top margins were producing 5-10 blank rows of empty space between visible blocks. Tool blocks now butt up against each other and the transcript reads denser and faster.

  20. v0.3.19

    2026-05-01

    Slash picker doesn't disappear when narrowing to one match

    Slash picker

    • ▸First pass at fixing the case where the picker disappeared when narrowing matches (typing `/m` then `c` made it vanish instead of showing `/mcp`). Improved further in v0.3.22 and finished in v0.3.23.

  21. v0.3.18

    2026-04-30

    Live token counts in the spinner + Fireworks cache hits

    Live spinner

    • ▸While the agent is thinking, the spinner now shows live ↑input / ↓output token counts that grow in real time as the response streams. No more staring at "Crunching... (3m 34s)" with zero feedback on how the request is progressing.

    Faster warm sessions

    • ▸Multi-turn conversations on DeepSeek V4 Pro now reuse prompt context across turns, so each follow-up turn returns its first token noticeably faster than the cold turn that preceded it. No config — kicks in automatically once the second turn fires.

  22. v0.3.17

    2026-04-30

    Read calls collapse into one bullet

    Transcript

    • ▸Six file reads in a row used to render six separate green-bullet rows. They now collapse into a single ● Searched (6) bullet with one branch per file — same UX as Grep/Glob/Ls already had. Less noise, easier to follow what the agent's looking at.
    • ▸Same grouping applies to Read, Grep, Glob, Ls, WebSearch, WebFetch, LSP, ListMemories, ReadMemory, ListSkills, ListAgents. Mutating tools (Edit, Write, Bash, NotebookEdit, Save*) stay per-call so diffs and command output are clearly visible.

  23. v0.3.16

    2026-04-30

    DeepSeek V4 Pro on Fireworks AI

    New model

    • ▸DeepSeek V4 Pro is now available — open /models and pick it like any other entry. Frontier reasoning + agentic tool use, 1M context window.
    • ▸Same monthly cap and top-up balance cover every model in the picker — switching across models doesn't double-bill you.

  24. v0.3.15

    2026-04-30

    Magic-link sign-in works again

    Sign-in

    • ▸Fixed the magic-link flow that was broken since v0.3.13 — every clicked sign-in email landed on a "server configuration" error. Magic-link sign-ins now complete normally on first click.
    • ▸Stale email links from the broken release still work — anyone with an unconsumed sign-in link in their inbox can use it without re-requesting.

  25. v0.3.14

    2026-04-29

    Tells you when you're over cap

    Over-cap UX

    • ▸When you're over your monthly token cap, fulcrum now tells you that directly: "Monthly cap exhausted (X / Y tokens on the Z tier). Top up at fulcrumcode.app/account, wait until the period resets, or upgrade tier." Previously the CLI just said your credentials were missing or expired — vague and misleading.
    • ▸Pre-flight detects over-cap state BEFORE prompting you to log in again. No more pointless browser dance that ends in an unexplained 429.

  26. v0.3.13

    2026-04-29

    Sign-in stays signed in

    Sign-in

    • ▸Fixed a loop where users were asked to sign in repeatedly during the same day, often within an hour of a fresh login. Auth and billing are now kept strictly separate — hitting your monthly cap no longer revokes your sign-in.
    • ▸Hitting your cap now shows a clear banner ("monthly cap full · top up at fulcrumcode.app/account · resets …") instead of forcing a re-login. Your sign-in stays valid.
    • ▸Credentials are now written to two places — the OS keychain and a local encrypted file. If keychain access breaks (binary update, OS keychain ACL change), the encrypted file backs you up so you don't have to log in again.

    Sign-in (cont'd)

    • ▸If something does go wrong with credential storage, the pre-flight check now tells you exactly which store failed instead of the previous vague "missing or expired" message.

  27. v0.3.12

    2026-04-29

    Faster start, working /whoami, cleaner sign-in

    Faster

    • ▸Cold start drops from ~10s to ~2s. The welcome banner paints immediately and any MCP servers connect in the background while you start typing.

    Sign-in

    • ▸If your sign-in expired or you never finished one, fulcrum prompts you to sign in BEFORE the TUI starts — not after, where you used to get stuck in a half-broken UI.
    • ▸`exit`, `:q`, and `/exit` all work in the rare case sign-in fails mid-session. Previously only Ctrl+C got you out.

    Bug fixes

    • ▸`/whoami` inside the TUI now shows your real cap, usage, and any unspent top-up balance. The 0.3.10 fix only landed in the standalone `fulcrum whoami` subcommand.

  28. v0.3.11

    2026-04-29

    VS Code extension is downloadable

    VS Code

    • ▸Install the Fulcrum extension with one command:
    • ▸`code --install-extension <(curl -L https://releases.fulcrumcode.app/latest/fulcrum-code-vscode.vsix -o /tmp/f.vsix && echo /tmp/f.vsix)`
    • ▸Or grab the .vsix from the Download page.

  29. v0.3.10

    2026-04-29

    /whoami shows the right numbers

    Bug fixes

    • ▸`/whoami` now shows your real monthly cap and usage. Top-up balance appears too if you have one.

  30. v0.3.9

    2026-04-29

    Banner correctly says 'signed in'

    Bug fixes

    • ▸Welcome banner showed 'not signed in' for some users whose session was valid. Now correct.

  31. v0.3.8

    2026-04-29

    Expired sign-in no longer crashes the TUI

    Bug fixes

    • ▸If your sign-in expires, fulcrum mounts cleanly with a 'run /login' hint instead of a traceback.

  32. v0.3.7

    2026-04-29

    VS Code extension preview.

    VS Code extension (preview)

    • ▸Activity-bar entry, send-selection, mention-active-file, sign-in. Install via .vsix (Marketplace listing in flight).

  33. v0.3.6

    2026-04-28

    Smaller download, faster startup, login keychain fix

    Faster + smaller

    • ▸macOS download is now 33 MB (down from 41 MB)
    • ▸Quick commands like `fulcrum --version` start ~2× faster

    Login

    • ▸Your sign-in token is now properly stored in the OS keychain on Linux and Windows (some installs were silently falling back to a local encrypted file)

  34. v0.3.5

    2026-04-28

    Login + update reliability

    Reliability

    • ▸Sign-in and `fulcrum update` now work on machines without a system certificate store (typical for fresh Linux containers)

  35. v0.3.4

    2026-04-28

    Update flow polish

    Reliability

    • ▸TLS verification fixed for sign-in and self-update on minimal Linux installs
    • ▸`fulcrum update` now reliably finds the latest version even on networks that block GitHub's API

  36. v0.3.3

    2026-04-28

    Self-updater, sign-in, and Claude-matched pricing

    Self-updater

    • ▸`fulcrum update` — one command to upgrade your installed binary to the latest release
    • ▸`fulcrum --check-update` — scriptable check that exits 1 if a newer version is available

    Pricing

    • ▸Tiers now match Claude — Starter $33, Pro $169, Team $339 (inc-GST AUD)
    • ▸Capacity shown as simple multipliers: 1× / 3× / 10×

    Sign-in

    • ▸Magic-link emails are reliable now — fixed a bug where the form silently failed and improved deliverability

  37. v0.3.2

    2026-04-27

    Tier 5 polish, marketplace, live-input UX

    Skill marketplace

    • ▸/marketplace add|list|search|install — clone any Claude Code-compatible marketplace.json repo (e.g. github.com/alirezarezvani/claude-skills with 235+ skills) and install plugins into ~/.fulcrum/skills/
    • ▸/skills lists every loaded skill with origin tag (bundled / claude / fulcrum)
    • ▸Recursive SKILL.md loader — installed plugins ship as <plugin>/<skill>/SKILL.md trees and fulcrum picks them up automatically

    Live-input UX

    • ▸Submit messages while the agent is running — they queue up and the agent works through them in order
    • ▸Conversational acknowledgement ("noted — I'll get to that as soon as I finish this turn") instead of robotic queue counters
    • ▸Persistent QueuePanel above the prompt shows everything pending; auto-hides when empty
    • ▸/cancel aborts the in-flight turn while keeping the queue intact
    • ▸/btw <question> — fork a side question off the current conversation; the answer streams to the transcript but never lands in the main history

    Transcript renderer

    • ▸Claude-Code-style ● Name(args) bullet per tool call
    • ▸Edit/Write tools show a real unified diff with line numbers + red/green backgrounds + "Added N lines, removed M lines" summary
    • ▸Bash shows $ command in cyan + dim description + truncated output with +N more lines
    • ▸Consecutive Grep/Glob/Ls calls collapse into one ● Searched bullet with one row per call
    • ▸Selectable transcript text + Cmd+C (macOS) / Ctrl+C (Linux/Windows) to copy

    Themes, vim mode, plugins, keys

    • ▸4 curated palettes (fulcrum-amber, fulcrum-midnight, fulcrum-mono, fulcrum-forest) on top of Textual's 20 built-ins; switchable at runtime via /theme
    • ▸Vim-style modal editing in the prompt (h/j/k/l, w/b, 0/$, gg/G, i/a/I/A/o/O, x, dd, yy, p, u) — toggle with /vim
    • ▸User plugins: drop a .py file in ~/.fulcrum/plugins/ to register new slash commands or tools
    • ▸Customisable keybindings via ~/.fulcrum/keybindings.json (open with /keys edit)

    Voice

    • ▸Animated "processing…" indicator with rotating glyph + dot cycle replaces the static "transcribing…" label

    Quality

    • ▸454 tests pass, 0 fail — first fully-green suite under any test ordering
    • ▸ToolSearch ranker now weights name matches above description matches so short, well-named tools beat tools whose descriptions happen to mention the query
    • ▸Save vision_model + transcription_model fields to ~/.fulcrum/config.json

  38. v0.2.2

    2026-04-25

    /voice + PyInstaller binary fixes

    Voice input

    • ▸/voice command records from the mic and transcribes with Whisper-Large-v3
    • ▸Inline recording indicator with elapsed timer, amplitude bar, and Space=stop / Esc=cancel

    Bundled binary

    • ▸Fixed tool-loading inside the PyInstaller-built executable so the bundled fulcrum CLI ships with the full toolset

  39. v0.2.1

    2026-04-25

    Bundled assets

    Bundled binary

    • ▸theme.css and skill data are now packaged into the PyInstaller binary so the standalone fulcrum executable runs without the source tree

  40. v0.2.0

    2026-04-25

    Tier 4 advanced tools

    New tools

    • ▸LSP — language-server-protocol bridge for IDE-quality diagnostics and definitions
    • ▸REPL — sandboxed Python evaluator the agent can drive between turns
    • ▸MCP — stdio + SSE bridges to any Model Context Protocol server
    • ▸Sub-agents — named profiles loaded from ~/.fulcrum/agents/*.md
    • ▸Cron — scheduled jobs the agent can register and trigger
    • ▸Notebook editing — direct Jupyter cell mutation
    • ▸Git worktrees — isolated branch checkouts for parallel work

    Slash commands

    • ▸20 built-in commands for session control: /config, /cost, /status, /compact, /history, /doctor, …

    Hooks

    • ▸User-configurable hooks for PreToolUse, UserPromptSubmit, SessionStart, and other lifecycle events — shell or Python

For a per-commit history, see the git log on GitHub.